xref: /freebsd/sys/cam/cam_periph.c (revision 118063fb70c3f74f16678dc7ddbc8528c483a528)
1 /*-
2  * Common functions for CAM "type" (peripheral) drivers.
3  *
4  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
5  *
6  * Copyright (c) 1997, 1998 Justin T. Gibbs.
7  * Copyright (c) 1997, 1998, 1999, 2000 Kenneth D. Merry.
8  * All rights reserved.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions, and the following disclaimer,
15  *    without modification, immediately at the beginning of the file.
16  * 2. The name of the author may not be used to endorse or promote products
17  *    derived from this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
23  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/types.h>
38 #include <sys/malloc.h>
39 #include <sys/kernel.h>
40 #include <sys/bio.h>
41 #include <sys/conf.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/buf.h>
45 #include <sys/proc.h>
46 #include <sys/devicestat.h>
47 #include <sys/bus.h>
48 #include <sys/sbuf.h>
49 #include <vm/vm.h>
50 #include <vm/vm_extern.h>
51 
52 #include <cam/cam.h>
53 #include <cam/cam_ccb.h>
54 #include <cam/cam_queue.h>
55 #include <cam/cam_xpt_periph.h>
56 #include <cam/cam_periph.h>
57 #include <cam/cam_debug.h>
58 #include <cam/cam_sim.h>
59 
60 #include <cam/scsi/scsi_all.h>
61 #include <cam/scsi/scsi_message.h>
62 #include <cam/scsi/scsi_pass.h>
63 
64 static	u_int		camperiphnextunit(struct periph_driver *p_drv,
65 					  u_int newunit, int wired,
66 					  path_id_t pathid, target_id_t target,
67 					  lun_id_t lun);
68 static	u_int		camperiphunit(struct periph_driver *p_drv,
69 				      path_id_t pathid, target_id_t target,
70 				      lun_id_t lun);
71 static	void		camperiphdone(struct cam_periph *periph,
72 					union ccb *done_ccb);
73 static  void		camperiphfree(struct cam_periph *periph);
74 static int		camperiphscsistatuserror(union ccb *ccb,
75 					        union ccb **orig_ccb,
76 						 cam_flags camflags,
77 						 u_int32_t sense_flags,
78 						 int *openings,
79 						 u_int32_t *relsim_flags,
80 						 u_int32_t *timeout,
81 						 u_int32_t  *action,
82 						 const char **action_string);
83 static	int		camperiphscsisenseerror(union ccb *ccb,
84 					        union ccb **orig_ccb,
85 					        cam_flags camflags,
86 					        u_int32_t sense_flags,
87 					        int *openings,
88 					        u_int32_t *relsim_flags,
89 					        u_int32_t *timeout,
90 					        u_int32_t *action,
91 					        const char **action_string);
92 static void		cam_periph_devctl_notify(union ccb *ccb);
93 
94 static int nperiph_drivers;
95 static int initialized = 0;
96 struct periph_driver **periph_drivers;
97 
98 static MALLOC_DEFINE(M_CAMPERIPH, "CAM periph", "CAM peripheral buffers");
99 
100 static int periph_selto_delay = 1000;
101 TUNABLE_INT("kern.cam.periph_selto_delay", &periph_selto_delay);
102 static int periph_noresrc_delay = 500;
103 TUNABLE_INT("kern.cam.periph_noresrc_delay", &periph_noresrc_delay);
104 static int periph_busy_delay = 500;
105 TUNABLE_INT("kern.cam.periph_busy_delay", &periph_busy_delay);
106 
107 
108 void
109 periphdriver_register(void *data)
110 {
111 	struct periph_driver *drv = (struct periph_driver *)data;
112 	struct periph_driver **newdrivers, **old;
113 	int ndrivers;
114 
115 again:
116 	ndrivers = nperiph_drivers + 2;
117 	newdrivers = malloc(sizeof(*newdrivers) * ndrivers, M_CAMPERIPH,
118 			    M_WAITOK);
119 	xpt_lock_buses();
120 	if (ndrivers != nperiph_drivers + 2) {
121 		/*
122 		 * Lost race against itself; go around.
123 		 */
124 		xpt_unlock_buses();
125 		free(newdrivers, M_CAMPERIPH);
126 		goto again;
127 	}
128 	if (periph_drivers)
129 		bcopy(periph_drivers, newdrivers,
130 		      sizeof(*newdrivers) * nperiph_drivers);
131 	newdrivers[nperiph_drivers] = drv;
132 	newdrivers[nperiph_drivers + 1] = NULL;
133 	old = periph_drivers;
134 	periph_drivers = newdrivers;
135 	nperiph_drivers++;
136 	xpt_unlock_buses();
137 	if (old)
138 		free(old, M_CAMPERIPH);
139 	/* If driver marked as early or it is late now, initialize it. */
140 	if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
141 	    initialized > 1)
142 		(*drv->init)();
143 }
144 
145 int
146 periphdriver_unregister(void *data)
147 {
148 	struct periph_driver *drv = (struct periph_driver *)data;
149 	int error, n;
150 
151 	/* If driver marked as early or it is late now, deinitialize it. */
152 	if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
153 	    initialized > 1) {
154 		if (drv->deinit == NULL) {
155 			printf("CAM periph driver '%s' doesn't have deinit.\n",
156 			    drv->driver_name);
157 			return (EOPNOTSUPP);
158 		}
159 		error = drv->deinit();
160 		if (error != 0)
161 			return (error);
162 	}
163 
164 	xpt_lock_buses();
165 	for (n = 0; n < nperiph_drivers && periph_drivers[n] != drv; n++)
166 		;
167 	KASSERT(n < nperiph_drivers,
168 	    ("Periph driver '%s' was not registered", drv->driver_name));
169 	for (; n + 1 < nperiph_drivers; n++)
170 		periph_drivers[n] = periph_drivers[n + 1];
171 	periph_drivers[n + 1] = NULL;
172 	nperiph_drivers--;
173 	xpt_unlock_buses();
174 	return (0);
175 }
176 
177 void
178 periphdriver_init(int level)
179 {
180 	int	i, early;
181 
182 	initialized = max(initialized, level);
183 	for (i = 0; periph_drivers[i] != NULL; i++) {
184 		early = (periph_drivers[i]->flags & CAM_PERIPH_DRV_EARLY) ? 1 : 2;
185 		if (early == initialized)
186 			(*periph_drivers[i]->init)();
187 	}
188 }
189 
190 cam_status
191 cam_periph_alloc(periph_ctor_t *periph_ctor,
192 		 periph_oninv_t *periph_oninvalidate,
193 		 periph_dtor_t *periph_dtor, periph_start_t *periph_start,
194 		 char *name, cam_periph_type type, struct cam_path *path,
195 		 ac_callback_t *ac_callback, ac_code code, void *arg)
196 {
197 	struct		periph_driver **p_drv;
198 	struct		cam_sim *sim;
199 	struct		cam_periph *periph;
200 	struct		cam_periph *cur_periph;
201 	path_id_t	path_id;
202 	target_id_t	target_id;
203 	lun_id_t	lun_id;
204 	cam_status	status;
205 	u_int		init_level;
206 
207 	init_level = 0;
208 	/*
209 	 * Handle Hot-Plug scenarios.  If there is already a peripheral
210 	 * of our type assigned to this path, we are likely waiting for
211 	 * final close on an old, invalidated, peripheral.  If this is
212 	 * the case, queue up a deferred call to the peripheral's async
213 	 * handler.  If it looks like a mistaken re-allocation, complain.
214 	 */
215 	if ((periph = cam_periph_find(path, name)) != NULL) {
216 
217 		if ((periph->flags & CAM_PERIPH_INVALID) != 0
218 		 && (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) == 0) {
219 			periph->flags |= CAM_PERIPH_NEW_DEV_FOUND;
220 			periph->deferred_callback = ac_callback;
221 			periph->deferred_ac = code;
222 			return (CAM_REQ_INPROG);
223 		} else {
224 			printf("cam_periph_alloc: attempt to re-allocate "
225 			       "valid device %s%d rejected flags %#x "
226 			       "refcount %d\n", periph->periph_name,
227 			       periph->unit_number, periph->flags,
228 			       periph->refcount);
229 		}
230 		return (CAM_REQ_INVALID);
231 	}
232 
233 	periph = (struct cam_periph *)malloc(sizeof(*periph), M_CAMPERIPH,
234 					     M_NOWAIT|M_ZERO);
235 
236 	if (periph == NULL)
237 		return (CAM_RESRC_UNAVAIL);
238 
239 	init_level++;
240 
241 
242 	sim = xpt_path_sim(path);
243 	path_id = xpt_path_path_id(path);
244 	target_id = xpt_path_target_id(path);
245 	lun_id = xpt_path_lun_id(path);
246 	periph->periph_start = periph_start;
247 	periph->periph_dtor = periph_dtor;
248 	periph->periph_oninval = periph_oninvalidate;
249 	periph->type = type;
250 	periph->periph_name = name;
251 	periph->scheduled_priority = CAM_PRIORITY_NONE;
252 	periph->immediate_priority = CAM_PRIORITY_NONE;
253 	periph->refcount = 1;		/* Dropped by invalidation. */
254 	periph->sim = sim;
255 	SLIST_INIT(&periph->ccb_list);
256 	status = xpt_create_path(&path, periph, path_id, target_id, lun_id);
257 	if (status != CAM_REQ_CMP)
258 		goto failure;
259 	periph->path = path;
260 
261 	xpt_lock_buses();
262 	for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
263 		if (strcmp((*p_drv)->driver_name, name) == 0)
264 			break;
265 	}
266 	if (*p_drv == NULL) {
267 		printf("cam_periph_alloc: invalid periph name '%s'\n", name);
268 		xpt_unlock_buses();
269 		xpt_free_path(periph->path);
270 		free(periph, M_CAMPERIPH);
271 		return (CAM_REQ_INVALID);
272 	}
273 	periph->unit_number = camperiphunit(*p_drv, path_id, target_id, lun_id);
274 	cur_periph = TAILQ_FIRST(&(*p_drv)->units);
275 	while (cur_periph != NULL
276 	    && cur_periph->unit_number < periph->unit_number)
277 		cur_periph = TAILQ_NEXT(cur_periph, unit_links);
278 	if (cur_periph != NULL) {
279 		KASSERT(cur_periph->unit_number != periph->unit_number, ("duplicate units on periph list"));
280 		TAILQ_INSERT_BEFORE(cur_periph, periph, unit_links);
281 	} else {
282 		TAILQ_INSERT_TAIL(&(*p_drv)->units, periph, unit_links);
283 		(*p_drv)->generation++;
284 	}
285 	xpt_unlock_buses();
286 
287 	init_level++;
288 
289 	status = xpt_add_periph(periph);
290 	if (status != CAM_REQ_CMP)
291 		goto failure;
292 
293 	init_level++;
294 	CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph created\n"));
295 
296 	status = periph_ctor(periph, arg);
297 
298 	if (status == CAM_REQ_CMP)
299 		init_level++;
300 
301 failure:
302 	switch (init_level) {
303 	case 4:
304 		/* Initialized successfully */
305 		break;
306 	case 3:
307 		CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
308 		xpt_remove_periph(periph);
309 		/* FALLTHROUGH */
310 	case 2:
311 		xpt_lock_buses();
312 		TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
313 		xpt_unlock_buses();
314 		xpt_free_path(periph->path);
315 		/* FALLTHROUGH */
316 	case 1:
317 		free(periph, M_CAMPERIPH);
318 		/* FALLTHROUGH */
319 	case 0:
320 		/* No cleanup to perform. */
321 		break;
322 	default:
323 		panic("%s: Unknown init level", __func__);
324 	}
325 	return(status);
326 }
327 
328 /*
329  * Find a peripheral structure with the specified path, target, lun,
330  * and (optionally) type.  If the name is NULL, this function will return
331  * the first peripheral driver that matches the specified path.
332  */
333 struct cam_periph *
334 cam_periph_find(struct cam_path *path, char *name)
335 {
336 	struct periph_driver **p_drv;
337 	struct cam_periph *periph;
338 
339 	xpt_lock_buses();
340 	for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
341 
342 		if (name != NULL && (strcmp((*p_drv)->driver_name, name) != 0))
343 			continue;
344 
345 		TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
346 			if (xpt_path_comp(periph->path, path) == 0) {
347 				xpt_unlock_buses();
348 				cam_periph_assert(periph, MA_OWNED);
349 				return(periph);
350 			}
351 		}
352 		if (name != NULL) {
353 			xpt_unlock_buses();
354 			return(NULL);
355 		}
356 	}
357 	xpt_unlock_buses();
358 	return(NULL);
359 }
360 
361 /*
362  * Find peripheral driver instances attached to the specified path.
363  */
364 int
365 cam_periph_list(struct cam_path *path, struct sbuf *sb)
366 {
367 	struct sbuf local_sb;
368 	struct periph_driver **p_drv;
369 	struct cam_periph *periph;
370 	int count;
371 	int sbuf_alloc_len;
372 
373 	sbuf_alloc_len = 16;
374 retry:
375 	sbuf_new(&local_sb, NULL, sbuf_alloc_len, SBUF_FIXEDLEN);
376 	count = 0;
377 	xpt_lock_buses();
378 	for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
379 
380 		TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
381 			if (xpt_path_comp(periph->path, path) != 0)
382 				continue;
383 
384 			if (sbuf_len(&local_sb) != 0)
385 				sbuf_cat(&local_sb, ",");
386 
387 			sbuf_printf(&local_sb, "%s%d", periph->periph_name,
388 				    periph->unit_number);
389 
390 			if (sbuf_error(&local_sb) == ENOMEM) {
391 				sbuf_alloc_len *= 2;
392 				xpt_unlock_buses();
393 				sbuf_delete(&local_sb);
394 				goto retry;
395 			}
396 			count++;
397 		}
398 	}
399 	xpt_unlock_buses();
400 	sbuf_finish(&local_sb);
401 	sbuf_cpy(sb, sbuf_data(&local_sb));
402 	sbuf_delete(&local_sb);
403 	return (count);
404 }
405 
406 int
407 cam_periph_acquire(struct cam_periph *periph)
408 {
409 	int status;
410 
411 	if (periph == NULL)
412 		return (EINVAL);
413 
414 	status = ENOENT;
415 	xpt_lock_buses();
416 	if ((periph->flags & CAM_PERIPH_INVALID) == 0) {
417 		periph->refcount++;
418 		status = 0;
419 	}
420 	xpt_unlock_buses();
421 
422 	return (status);
423 }
424 
425 void
426 cam_periph_doacquire(struct cam_periph *periph)
427 {
428 
429 	xpt_lock_buses();
430 	KASSERT(periph->refcount >= 1,
431 	    ("cam_periph_doacquire() with refcount == %d", periph->refcount));
432 	periph->refcount++;
433 	xpt_unlock_buses();
434 }
435 
436 void
437 cam_periph_release_locked_buses(struct cam_periph *periph)
438 {
439 
440 	cam_periph_assert(periph, MA_OWNED);
441 	KASSERT(periph->refcount >= 1, ("periph->refcount >= 1"));
442 	if (--periph->refcount == 0)
443 		camperiphfree(periph);
444 }
445 
446 void
447 cam_periph_release_locked(struct cam_periph *periph)
448 {
449 
450 	if (periph == NULL)
451 		return;
452 
453 	xpt_lock_buses();
454 	cam_periph_release_locked_buses(periph);
455 	xpt_unlock_buses();
456 }
457 
458 void
459 cam_periph_release(struct cam_periph *periph)
460 {
461 	struct mtx *mtx;
462 
463 	if (periph == NULL)
464 		return;
465 
466 	cam_periph_assert(periph, MA_NOTOWNED);
467 	mtx = cam_periph_mtx(periph);
468 	mtx_lock(mtx);
469 	cam_periph_release_locked(periph);
470 	mtx_unlock(mtx);
471 }
472 
473 int
474 cam_periph_hold(struct cam_periph *periph, int priority)
475 {
476 	int error;
477 
478 	/*
479 	 * Increment the reference count on the peripheral
480 	 * while we wait for our lock attempt to succeed
481 	 * to ensure the peripheral doesn't disappear out
482 	 * from user us while we sleep.
483 	 */
484 
485 	if (cam_periph_acquire(periph) != 0)
486 		return (ENXIO);
487 
488 	cam_periph_assert(periph, MA_OWNED);
489 	while ((periph->flags & CAM_PERIPH_LOCKED) != 0) {
490 		periph->flags |= CAM_PERIPH_LOCK_WANTED;
491 		if ((error = cam_periph_sleep(periph, periph, priority,
492 		    "caplck", 0)) != 0) {
493 			cam_periph_release_locked(periph);
494 			return (error);
495 		}
496 		if (periph->flags & CAM_PERIPH_INVALID) {
497 			cam_periph_release_locked(periph);
498 			return (ENXIO);
499 		}
500 	}
501 
502 	periph->flags |= CAM_PERIPH_LOCKED;
503 	return (0);
504 }
505 
506 void
507 cam_periph_unhold(struct cam_periph *periph)
508 {
509 
510 	cam_periph_assert(periph, MA_OWNED);
511 
512 	periph->flags &= ~CAM_PERIPH_LOCKED;
513 	if ((periph->flags & CAM_PERIPH_LOCK_WANTED) != 0) {
514 		periph->flags &= ~CAM_PERIPH_LOCK_WANTED;
515 		wakeup(periph);
516 	}
517 
518 	cam_periph_release_locked(periph);
519 }
520 
521 /*
522  * Look for the next unit number that is not currently in use for this
523  * peripheral type starting at "newunit".  Also exclude unit numbers that
524  * are reserved by for future "hardwiring" unless we already know that this
525  * is a potential wired device.  Only assume that the device is "wired" the
526  * first time through the loop since after that we'll be looking at unit
527  * numbers that did not match a wiring entry.
528  */
529 static u_int
530 camperiphnextunit(struct periph_driver *p_drv, u_int newunit, int wired,
531 		  path_id_t pathid, target_id_t target, lun_id_t lun)
532 {
533 	struct	cam_periph *periph;
534 	char	*periph_name;
535 	int	i, val, dunit, r;
536 	const char *dname, *strval;
537 
538 	periph_name = p_drv->driver_name;
539 	for (;;newunit++) {
540 
541 		for (periph = TAILQ_FIRST(&p_drv->units);
542 		     periph != NULL && periph->unit_number != newunit;
543 		     periph = TAILQ_NEXT(periph, unit_links))
544 			;
545 
546 		if (periph != NULL && periph->unit_number == newunit) {
547 			if (wired != 0) {
548 				xpt_print(periph->path, "Duplicate Wired "
549 				    "Device entry!\n");
550 				xpt_print(periph->path, "Second device (%s "
551 				    "device at scbus%d target %d lun %d) will "
552 				    "not be wired\n", periph_name, pathid,
553 				    target, lun);
554 				wired = 0;
555 			}
556 			continue;
557 		}
558 		if (wired)
559 			break;
560 
561 		/*
562 		 * Don't match entries like "da 4" as a wired down
563 		 * device, but do match entries like "da 4 target 5"
564 		 * or even "da 4 scbus 1".
565 		 */
566 		i = 0;
567 		dname = periph_name;
568 		for (;;) {
569 			r = resource_find_dev(&i, dname, &dunit, NULL, NULL);
570 			if (r != 0)
571 				break;
572 			/* if no "target" and no specific scbus, skip */
573 			if (resource_int_value(dname, dunit, "target", &val) &&
574 			    (resource_string_value(dname, dunit, "at",&strval)||
575 			     strcmp(strval, "scbus") == 0))
576 				continue;
577 			if (newunit == dunit)
578 				break;
579 		}
580 		if (r != 0)
581 			break;
582 	}
583 	return (newunit);
584 }
585 
586 static u_int
587 camperiphunit(struct periph_driver *p_drv, path_id_t pathid,
588 	      target_id_t target, lun_id_t lun)
589 {
590 	u_int	unit;
591 	int	wired, i, val, dunit;
592 	const char *dname, *strval;
593 	char	pathbuf[32], *periph_name;
594 
595 	periph_name = p_drv->driver_name;
596 	snprintf(pathbuf, sizeof(pathbuf), "scbus%d", pathid);
597 	unit = 0;
598 	i = 0;
599 	dname = periph_name;
600 	for (wired = 0; resource_find_dev(&i, dname, &dunit, NULL, NULL) == 0;
601 	     wired = 0) {
602 		if (resource_string_value(dname, dunit, "at", &strval) == 0) {
603 			if (strcmp(strval, pathbuf) != 0)
604 				continue;
605 			wired++;
606 		}
607 		if (resource_int_value(dname, dunit, "target", &val) == 0) {
608 			if (val != target)
609 				continue;
610 			wired++;
611 		}
612 		if (resource_int_value(dname, dunit, "lun", &val) == 0) {
613 			if (val != lun)
614 				continue;
615 			wired++;
616 		}
617 		if (wired != 0) {
618 			unit = dunit;
619 			break;
620 		}
621 	}
622 
623 	/*
624 	 * Either start from 0 looking for the next unit or from
625 	 * the unit number given in the resource config.  This way,
626 	 * if we have wildcard matches, we don't return the same
627 	 * unit number twice.
628 	 */
629 	unit = camperiphnextunit(p_drv, unit, wired, pathid, target, lun);
630 
631 	return (unit);
632 }
633 
634 void
635 cam_periph_invalidate(struct cam_periph *periph)
636 {
637 
638 	cam_periph_assert(periph, MA_OWNED);
639 	/*
640 	 * We only call this routine the first time a peripheral is
641 	 * invalidated.
642 	 */
643 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
644 		return;
645 
646 	CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph invalidated\n"));
647 	if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting) {
648 		struct sbuf sb;
649 		char buffer[160];
650 
651 		sbuf_new(&sb, buffer, 160, SBUF_FIXEDLEN);
652 		xpt_denounce_periph_sbuf(periph, &sb);
653 		sbuf_finish(&sb);
654 		sbuf_putbuf(&sb);
655 	}
656 	periph->flags |= CAM_PERIPH_INVALID;
657 	periph->flags &= ~CAM_PERIPH_NEW_DEV_FOUND;
658 	if (periph->periph_oninval != NULL)
659 		periph->periph_oninval(periph);
660 	cam_periph_release_locked(periph);
661 }
662 
663 static void
664 camperiphfree(struct cam_periph *periph)
665 {
666 	struct periph_driver **p_drv;
667 	struct periph_driver *drv;
668 
669 	cam_periph_assert(periph, MA_OWNED);
670 	KASSERT(periph->periph_allocating == 0, ("%s%d: freed while allocating",
671 	    periph->periph_name, periph->unit_number));
672 	for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
673 		if (strcmp((*p_drv)->driver_name, periph->periph_name) == 0)
674 			break;
675 	}
676 	if (*p_drv == NULL) {
677 		printf("camperiphfree: attempt to free non-existant periph\n");
678 		return;
679 	}
680 	/*
681 	 * Cache a pointer to the periph_driver structure.  If a
682 	 * periph_driver is added or removed from the array (see
683 	 * periphdriver_register()) while we drop the toplogy lock
684 	 * below, p_drv may change.  This doesn't protect against this
685 	 * particular periph_driver going away.  That will require full
686 	 * reference counting in the periph_driver infrastructure.
687 	 */
688 	drv = *p_drv;
689 
690 	/*
691 	 * We need to set this flag before dropping the topology lock, to
692 	 * let anyone who is traversing the list that this peripheral is
693 	 * about to be freed, and there will be no more reference count
694 	 * checks.
695 	 */
696 	periph->flags |= CAM_PERIPH_FREE;
697 
698 	/*
699 	 * The peripheral destructor semantics dictate calling with only the
700 	 * SIM mutex held.  Since it might sleep, it should not be called
701 	 * with the topology lock held.
702 	 */
703 	xpt_unlock_buses();
704 
705 	/*
706 	 * We need to call the peripheral destructor prior to removing the
707 	 * peripheral from the list.  Otherwise, we risk running into a
708 	 * scenario where the peripheral unit number may get reused
709 	 * (because it has been removed from the list), but some resources
710 	 * used by the peripheral are still hanging around.  In particular,
711 	 * the devfs nodes used by some peripherals like the pass(4) driver
712 	 * aren't fully cleaned up until the destructor is run.  If the
713 	 * unit number is reused before the devfs instance is fully gone,
714 	 * devfs will panic.
715 	 */
716 	if (periph->periph_dtor != NULL)
717 		periph->periph_dtor(periph);
718 
719 	/*
720 	 * The peripheral list is protected by the topology lock.
721 	 */
722 	xpt_lock_buses();
723 
724 	TAILQ_REMOVE(&drv->units, periph, unit_links);
725 	drv->generation++;
726 
727 	xpt_remove_periph(periph);
728 
729 	xpt_unlock_buses();
730 	if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting)
731 		xpt_print(periph->path, "Periph destroyed\n");
732 	else
733 		CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
734 
735 	if (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) {
736 		union ccb ccb;
737 		void *arg;
738 
739 		switch (periph->deferred_ac) {
740 		case AC_FOUND_DEVICE:
741 			ccb.ccb_h.func_code = XPT_GDEV_TYPE;
742 			xpt_setup_ccb(&ccb.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
743 			xpt_action(&ccb);
744 			arg = &ccb;
745 			break;
746 		case AC_PATH_REGISTERED:
747 			xpt_path_inq(&ccb.cpi, periph->path);
748 			arg = &ccb;
749 			break;
750 		default:
751 			arg = NULL;
752 			break;
753 		}
754 		periph->deferred_callback(NULL, periph->deferred_ac,
755 					  periph->path, arg);
756 	}
757 	xpt_free_path(periph->path);
758 	free(periph, M_CAMPERIPH);
759 	xpt_lock_buses();
760 }
761 
762 /*
763  * Map user virtual pointers into kernel virtual address space, so we can
764  * access the memory.  This is now a generic function that centralizes most
765  * of the sanity checks on the data flags, if any.
766  * This also only works for up to MAXPHYS memory.  Since we use
767  * buffers to map stuff in and out, we're limited to the buffer size.
768  */
769 int
770 cam_periph_mapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo,
771     u_int maxmap)
772 {
773 	int numbufs, i, j;
774 	int flags[CAM_PERIPH_MAXMAPS];
775 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
776 	u_int32_t lengths[CAM_PERIPH_MAXMAPS];
777 	u_int32_t dirs[CAM_PERIPH_MAXMAPS];
778 
779 	if (maxmap == 0)
780 		maxmap = DFLTPHYS;	/* traditional default */
781 	else if (maxmap > MAXPHYS)
782 		maxmap = MAXPHYS;	/* for safety */
783 	switch(ccb->ccb_h.func_code) {
784 	case XPT_DEV_MATCH:
785 		if (ccb->cdm.match_buf_len == 0) {
786 			printf("cam_periph_mapmem: invalid match buffer "
787 			       "length 0\n");
788 			return(EINVAL);
789 		}
790 		if (ccb->cdm.pattern_buf_len > 0) {
791 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
792 			lengths[0] = ccb->cdm.pattern_buf_len;
793 			dirs[0] = CAM_DIR_OUT;
794 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
795 			lengths[1] = ccb->cdm.match_buf_len;
796 			dirs[1] = CAM_DIR_IN;
797 			numbufs = 2;
798 		} else {
799 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
800 			lengths[0] = ccb->cdm.match_buf_len;
801 			dirs[0] = CAM_DIR_IN;
802 			numbufs = 1;
803 		}
804 		/*
805 		 * This request will not go to the hardware, no reason
806 		 * to be so strict. vmapbuf() is able to map up to MAXPHYS.
807 		 */
808 		maxmap = MAXPHYS;
809 		break;
810 	case XPT_SCSI_IO:
811 	case XPT_CONT_TARGET_IO:
812 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
813 			return(0);
814 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
815 			return (EINVAL);
816 		data_ptrs[0] = &ccb->csio.data_ptr;
817 		lengths[0] = ccb->csio.dxfer_len;
818 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
819 		numbufs = 1;
820 		break;
821 	case XPT_ATA_IO:
822 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
823 			return(0);
824 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
825 			return (EINVAL);
826 		data_ptrs[0] = &ccb->ataio.data_ptr;
827 		lengths[0] = ccb->ataio.dxfer_len;
828 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
829 		numbufs = 1;
830 		break;
831 	case XPT_MMC_IO:
832 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
833 			return(0);
834 		/* Two mappings: one for cmd->data and one for cmd->data->data */
835 		data_ptrs[0] = (unsigned char **)&ccb->mmcio.cmd.data;
836 		lengths[0] = sizeof(struct mmc_data *);
837 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
838 		data_ptrs[1] = (unsigned char **)&ccb->mmcio.cmd.data->data;
839 		lengths[1] = ccb->mmcio.cmd.data->len;
840 		dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
841 		numbufs = 2;
842 		break;
843 	case XPT_SMP_IO:
844 		data_ptrs[0] = &ccb->smpio.smp_request;
845 		lengths[0] = ccb->smpio.smp_request_len;
846 		dirs[0] = CAM_DIR_OUT;
847 		data_ptrs[1] = &ccb->smpio.smp_response;
848 		lengths[1] = ccb->smpio.smp_response_len;
849 		dirs[1] = CAM_DIR_IN;
850 		numbufs = 2;
851 		break;
852 	case XPT_NVME_IO:
853 	case XPT_NVME_ADMIN:
854 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
855 			return (0);
856 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
857 			return (EINVAL);
858 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
859 		lengths[0] = ccb->nvmeio.dxfer_len;
860 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
861 		numbufs = 1;
862 		break;
863 	case XPT_DEV_ADVINFO:
864 		if (ccb->cdai.bufsiz == 0)
865 			return (0);
866 
867 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
868 		lengths[0] = ccb->cdai.bufsiz;
869 		dirs[0] = CAM_DIR_IN;
870 		numbufs = 1;
871 
872 		/*
873 		 * This request will not go to the hardware, no reason
874 		 * to be so strict. vmapbuf() is able to map up to MAXPHYS.
875 		 */
876 		maxmap = MAXPHYS;
877 		break;
878 	default:
879 		return(EINVAL);
880 		break; /* NOTREACHED */
881 	}
882 
883 	/*
884 	 * Check the transfer length and permissions first, so we don't
885 	 * have to unmap any previously mapped buffers.
886 	 */
887 	for (i = 0; i < numbufs; i++) {
888 
889 		flags[i] = 0;
890 
891 		/*
892 		 * The userland data pointer passed in may not be page
893 		 * aligned.  vmapbuf() truncates the address to a page
894 		 * boundary, so if the address isn't page aligned, we'll
895 		 * need enough space for the given transfer length, plus
896 		 * whatever extra space is necessary to make it to the page
897 		 * boundary.
898 		 */
899 		if ((lengths[i] +
900 		    (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)) > maxmap){
901 			printf("cam_periph_mapmem: attempt to map %lu bytes, "
902 			       "which is greater than %lu\n",
903 			       (long)(lengths[i] +
904 			       (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)),
905 			       (u_long)maxmap);
906 			return(E2BIG);
907 		}
908 
909 		if (dirs[i] & CAM_DIR_OUT) {
910 			flags[i] = BIO_WRITE;
911 		}
912 
913 		if (dirs[i] & CAM_DIR_IN) {
914 			flags[i] = BIO_READ;
915 		}
916 
917 	}
918 
919 	/*
920 	 * This keeps the kernel stack of current thread from getting
921 	 * swapped.  In low-memory situations where the kernel stack might
922 	 * otherwise get swapped out, this holds it and allows the thread
923 	 * to make progress and release the kernel mapped pages sooner.
924 	 *
925 	 * XXX KDM should I use P_NOSWAP instead?
926 	 */
927 	PHOLD(curproc);
928 
929 	for (i = 0; i < numbufs; i++) {
930 		/*
931 		 * Get the buffer.
932 		 */
933 		mapinfo->bp[i] = getpbuf(NULL);
934 
935 		/* put our pointer in the data slot */
936 		mapinfo->bp[i]->b_data = *data_ptrs[i];
937 
938 		/* save the user's data address */
939 		mapinfo->bp[i]->b_caller1 = *data_ptrs[i];
940 
941 		/* set the transfer length, we know it's < MAXPHYS */
942 		mapinfo->bp[i]->b_bufsize = lengths[i];
943 
944 		/* set the direction */
945 		mapinfo->bp[i]->b_iocmd = flags[i];
946 
947 		/*
948 		 * Map the buffer into kernel memory.
949 		 *
950 		 * Note that useracc() alone is not a  sufficient test.
951 		 * vmapbuf() can still fail due to a smaller file mapped
952 		 * into a larger area of VM, or if userland races against
953 		 * vmapbuf() after the useracc() check.
954 		 */
955 		if (vmapbuf(mapinfo->bp[i], 1) < 0) {
956 			for (j = 0; j < i; ++j) {
957 				*data_ptrs[j] = mapinfo->bp[j]->b_caller1;
958 				vunmapbuf(mapinfo->bp[j]);
959 				relpbuf(mapinfo->bp[j], NULL);
960 			}
961 			relpbuf(mapinfo->bp[i], NULL);
962 			PRELE(curproc);
963 			return(EACCES);
964 		}
965 
966 		/* set our pointer to the new mapped area */
967 		*data_ptrs[i] = mapinfo->bp[i]->b_data;
968 
969 		mapinfo->num_bufs_used++;
970 	}
971 
972 	/*
973 	 * Now that we've gotten this far, change ownership to the kernel
974 	 * of the buffers so that we don't run afoul of returning to user
975 	 * space with locks (on the buffer) held.
976 	 */
977 	for (i = 0; i < numbufs; i++) {
978 		BUF_KERNPROC(mapinfo->bp[i]);
979 	}
980 
981 
982 	return(0);
983 }
984 
985 /*
986  * Unmap memory segments mapped into kernel virtual address space by
987  * cam_periph_mapmem().
988  */
989 void
990 cam_periph_unmapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
991 {
992 	int numbufs, i;
993 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
994 
995 	if (mapinfo->num_bufs_used <= 0) {
996 		/* nothing to free and the process wasn't held. */
997 		return;
998 	}
999 
1000 	switch (ccb->ccb_h.func_code) {
1001 	case XPT_DEV_MATCH:
1002 		numbufs = min(mapinfo->num_bufs_used, 2);
1003 
1004 		if (numbufs == 1) {
1005 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1006 		} else {
1007 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1008 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1009 		}
1010 		break;
1011 	case XPT_SCSI_IO:
1012 	case XPT_CONT_TARGET_IO:
1013 		data_ptrs[0] = &ccb->csio.data_ptr;
1014 		numbufs = min(mapinfo->num_bufs_used, 1);
1015 		break;
1016 	case XPT_ATA_IO:
1017 		data_ptrs[0] = &ccb->ataio.data_ptr;
1018 		numbufs = min(mapinfo->num_bufs_used, 1);
1019 		break;
1020 	case XPT_SMP_IO:
1021 		numbufs = min(mapinfo->num_bufs_used, 2);
1022 		data_ptrs[0] = &ccb->smpio.smp_request;
1023 		data_ptrs[1] = &ccb->smpio.smp_response;
1024 		break;
1025 	case XPT_DEV_ADVINFO:
1026 		numbufs = min(mapinfo->num_bufs_used, 1);
1027 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1028 		break;
1029 	case XPT_NVME_IO:
1030 	case XPT_NVME_ADMIN:
1031 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1032 		numbufs = min(mapinfo->num_bufs_used, 1);
1033 		break;
1034 	default:
1035 		/* allow ourselves to be swapped once again */
1036 		PRELE(curproc);
1037 		return;
1038 		break; /* NOTREACHED */
1039 	}
1040 
1041 	for (i = 0; i < numbufs; i++) {
1042 		/* Set the user's pointer back to the original value */
1043 		*data_ptrs[i] = mapinfo->bp[i]->b_caller1;
1044 
1045 		/* unmap the buffer */
1046 		vunmapbuf(mapinfo->bp[i]);
1047 
1048 		/* release the buffer */
1049 		relpbuf(mapinfo->bp[i], NULL);
1050 	}
1051 
1052 	/* allow ourselves to be swapped once again */
1053 	PRELE(curproc);
1054 }
1055 
1056 int
1057 cam_periph_ioctl(struct cam_periph *periph, u_long cmd, caddr_t addr,
1058 		 int (*error_routine)(union ccb *ccb,
1059 				      cam_flags camflags,
1060 				      u_int32_t sense_flags))
1061 {
1062 	union ccb 	     *ccb;
1063 	int 		     error;
1064 	int		     found;
1065 
1066 	error = found = 0;
1067 
1068 	switch(cmd){
1069 	case CAMGETPASSTHRU:
1070 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1071 		xpt_setup_ccb(&ccb->ccb_h,
1072 			      ccb->ccb_h.path,
1073 			      CAM_PRIORITY_NORMAL);
1074 		ccb->ccb_h.func_code = XPT_GDEVLIST;
1075 
1076 		/*
1077 		 * Basically, the point of this is that we go through
1078 		 * getting the list of devices, until we find a passthrough
1079 		 * device.  In the current version of the CAM code, the
1080 		 * only way to determine what type of device we're dealing
1081 		 * with is by its name.
1082 		 */
1083 		while (found == 0) {
1084 			ccb->cgdl.index = 0;
1085 			ccb->cgdl.status = CAM_GDEVLIST_MORE_DEVS;
1086 			while (ccb->cgdl.status == CAM_GDEVLIST_MORE_DEVS) {
1087 
1088 				/* we want the next device in the list */
1089 				xpt_action(ccb);
1090 				if (strncmp(ccb->cgdl.periph_name,
1091 				    "pass", 4) == 0){
1092 					found = 1;
1093 					break;
1094 				}
1095 			}
1096 			if ((ccb->cgdl.status == CAM_GDEVLIST_LAST_DEVICE) &&
1097 			    (found == 0)) {
1098 				ccb->cgdl.periph_name[0] = '\0';
1099 				ccb->cgdl.unit_number = 0;
1100 				break;
1101 			}
1102 		}
1103 
1104 		/* copy the result back out */
1105 		bcopy(ccb, addr, sizeof(union ccb));
1106 
1107 		/* and release the ccb */
1108 		xpt_release_ccb(ccb);
1109 
1110 		break;
1111 	default:
1112 		error = ENOTTY;
1113 		break;
1114 	}
1115 	return(error);
1116 }
1117 
1118 static void
1119 cam_periph_done_panic(struct cam_periph *periph, union ccb *done_ccb)
1120 {
1121 
1122 	panic("%s: already done with ccb %p", __func__, done_ccb);
1123 }
1124 
1125 static void
1126 cam_periph_done(struct cam_periph *periph, union ccb *done_ccb)
1127 {
1128 
1129 	/* Caller will release the CCB */
1130 	xpt_path_assert(done_ccb->ccb_h.path, MA_OWNED);
1131 	done_ccb->ccb_h.cbfcnp = cam_periph_done_panic;
1132 	wakeup(&done_ccb->ccb_h.cbfcnp);
1133 }
1134 
1135 static void
1136 cam_periph_ccbwait(union ccb *ccb)
1137 {
1138 
1139 	if ((ccb->ccb_h.func_code & XPT_FC_QUEUED) != 0) {
1140 		while (ccb->ccb_h.cbfcnp != cam_periph_done_panic)
1141 			xpt_path_sleep(ccb->ccb_h.path, &ccb->ccb_h.cbfcnp,
1142 			    PRIBIO, "cbwait", 0);
1143 	}
1144 	KASSERT(ccb->ccb_h.pinfo.index == CAM_UNQUEUED_INDEX &&
1145 	    (ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_INPROG,
1146 	    ("%s: proceeding with incomplete ccb: ccb=%p, func_code=%#x, "
1147 	     "status=%#x, index=%d", __func__, ccb, ccb->ccb_h.func_code,
1148 	     ccb->ccb_h.status, ccb->ccb_h.pinfo.index));
1149 }
1150 
1151 int
1152 cam_periph_runccb(union ccb *ccb,
1153 		  int (*error_routine)(union ccb *ccb,
1154 				       cam_flags camflags,
1155 				       u_int32_t sense_flags),
1156 		  cam_flags camflags, u_int32_t sense_flags,
1157 		  struct devstat *ds)
1158 {
1159 	struct bintime *starttime;
1160 	struct bintime ltime;
1161 	int error;
1162 	bool must_poll;
1163 	uint32_t timeout = 1;
1164 
1165 	starttime = NULL;
1166 	xpt_path_assert(ccb->ccb_h.path, MA_OWNED);
1167 	KASSERT((ccb->ccb_h.flags & CAM_UNLOCKED) == 0,
1168 	    ("%s: ccb=%p, func_code=%#x, flags=%#x", __func__, ccb,
1169 	     ccb->ccb_h.func_code, ccb->ccb_h.flags));
1170 
1171 	/*
1172 	 * If the user has supplied a stats structure, and if we understand
1173 	 * this particular type of ccb, record the transaction start.
1174 	 */
1175 	if (ds != NULL &&
1176 	    (ccb->ccb_h.func_code == XPT_SCSI_IO ||
1177 	    ccb->ccb_h.func_code == XPT_ATA_IO ||
1178 	    ccb->ccb_h.func_code == XPT_NVME_IO)) {
1179 		starttime = &ltime;
1180 		binuptime(starttime);
1181 		devstat_start_transaction(ds, starttime);
1182 	}
1183 
1184 	/*
1185 	 * We must poll the I/O while we're dumping. The scheduler is normally
1186 	 * stopped for dumping, except when we call doadump from ddb. While the
1187 	 * scheduler is running in this case, we still need to poll the I/O to
1188 	 * avoid sleeping waiting for the ccb to complete.
1189 	 *
1190 	 * A panic triggered dump stops the scheduler, any callback from the
1191 	 * shutdown_post_sync event will run with the scheduler stopped, but
1192 	 * before we're officially dumping. To avoid hanging in adashutdown
1193 	 * initiated commands (or other similar situations), we have to test for
1194 	 * either SCHEDULER_STOPPED() here as well.
1195 	 *
1196 	 * To avoid locking problems, dumping/polling callers must call
1197 	 * without a periph lock held.
1198 	 */
1199 	must_poll = dumping || SCHEDULER_STOPPED();
1200 	ccb->ccb_h.cbfcnp = cam_periph_done;
1201 
1202 	/*
1203 	 * If we're polling, then we need to ensure that we have ample resources
1204 	 * in the periph.
1205 	 * cam_periph_error can reschedule the ccb by calling xpt_action and returning
1206 	 * ERESTART, so we have to effect the polling in the do loop below.
1207 	 */
1208 	if (must_poll) {
1209 		timeout = xpt_poll_setup(ccb);
1210 	}
1211 
1212 	if (timeout == 0) {
1213 		ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
1214 		error = EBUSY;
1215 	} else {
1216 		xpt_action(ccb);
1217 		do {
1218 			if (must_poll) {
1219 				xpt_pollwait(ccb, timeout);
1220 				timeout = ccb->ccb_h.timeout * 10;
1221 			} else {
1222 				cam_periph_ccbwait(ccb);
1223 			}
1224 			if ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP)
1225 				error = 0;
1226 			else if (error_routine != NULL) {
1227 				ccb->ccb_h.cbfcnp = cam_periph_done;
1228 				error = (*error_routine)(ccb, camflags, sense_flags);
1229 			} else
1230 				error = 0;
1231 		} while (error == ERESTART);
1232 	}
1233 
1234 	if ((ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) {
1235 		cam_release_devq(ccb->ccb_h.path,
1236 				 /* relsim_flags */0,
1237 				 /* openings */0,
1238 				 /* timeout */0,
1239 				 /* getcount_only */ FALSE);
1240 		ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1241 	}
1242 
1243 	if (ds != NULL) {
1244 		uint32_t bytes;
1245 		devstat_tag_type tag;
1246 		bool valid = true;
1247 
1248 		if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
1249 			bytes = ccb->csio.dxfer_len - ccb->csio.resid;
1250 			tag = (devstat_tag_type)(ccb->csio.tag_action & 0x3);
1251 		} else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
1252 			bytes = ccb->ataio.dxfer_len - ccb->ataio.resid;
1253 			tag = (devstat_tag_type)0;
1254 		} else if (ccb->ccb_h.func_code == XPT_NVME_IO) {
1255 			bytes = ccb->nvmeio.dxfer_len; /* NB: resid no possible */
1256 			tag = (devstat_tag_type)0;
1257 		} else {
1258 			valid = false;
1259 		}
1260 		if (valid)
1261 			devstat_end_transaction(ds, bytes, tag,
1262 			    ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE) ?
1263 			    DEVSTAT_NO_DATA : (ccb->ccb_h.flags & CAM_DIR_OUT) ?
1264 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL, starttime);
1265 	}
1266 
1267 	return(error);
1268 }
1269 
1270 void
1271 cam_freeze_devq(struct cam_path *path)
1272 {
1273 	struct ccb_hdr ccb_h;
1274 
1275 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_freeze_devq\n"));
1276 	xpt_setup_ccb(&ccb_h, path, /*priority*/1);
1277 	ccb_h.func_code = XPT_NOOP;
1278 	ccb_h.flags = CAM_DEV_QFREEZE;
1279 	xpt_action((union ccb *)&ccb_h);
1280 }
1281 
1282 u_int32_t
1283 cam_release_devq(struct cam_path *path, u_int32_t relsim_flags,
1284 		 u_int32_t openings, u_int32_t arg,
1285 		 int getcount_only)
1286 {
1287 	struct ccb_relsim crs;
1288 
1289 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_release_devq(%u, %u, %u, %d)\n",
1290 	    relsim_flags, openings, arg, getcount_only));
1291 	xpt_setup_ccb(&crs.ccb_h, path, CAM_PRIORITY_NORMAL);
1292 	crs.ccb_h.func_code = XPT_REL_SIMQ;
1293 	crs.ccb_h.flags = getcount_only ? CAM_DEV_QFREEZE : 0;
1294 	crs.release_flags = relsim_flags;
1295 	crs.openings = openings;
1296 	crs.release_timeout = arg;
1297 	xpt_action((union ccb *)&crs);
1298 	return (crs.qfrozen_cnt);
1299 }
1300 
1301 #define saved_ccb_ptr ppriv_ptr0
1302 static void
1303 camperiphdone(struct cam_periph *periph, union ccb *done_ccb)
1304 {
1305 	union ccb      *saved_ccb;
1306 	cam_status	status;
1307 	struct scsi_start_stop_unit *scsi_cmd;
1308 	int    error_code, sense_key, asc, ascq;
1309 
1310 	scsi_cmd = (struct scsi_start_stop_unit *)
1311 	    &done_ccb->csio.cdb_io.cdb_bytes;
1312 	status = done_ccb->ccb_h.status;
1313 
1314 	if ((status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1315 		if (scsi_extract_sense_ccb(done_ccb,
1316 		    &error_code, &sense_key, &asc, &ascq)) {
1317 			/*
1318 			 * If the error is "invalid field in CDB",
1319 			 * and the load/eject flag is set, turn the
1320 			 * flag off and try again.  This is just in
1321 			 * case the drive in question barfs on the
1322 			 * load eject flag.  The CAM code should set
1323 			 * the load/eject flag by default for
1324 			 * removable media.
1325 			 */
1326 			if ((scsi_cmd->opcode == START_STOP_UNIT) &&
1327 			    ((scsi_cmd->how & SSS_LOEJ) != 0) &&
1328 			     (asc == 0x24) && (ascq == 0x00)) {
1329 				scsi_cmd->how &= ~SSS_LOEJ;
1330 				if (status & CAM_DEV_QFRZN) {
1331 					cam_release_devq(done_ccb->ccb_h.path,
1332 					    0, 0, 0, 0);
1333 					done_ccb->ccb_h.status &=
1334 					    ~CAM_DEV_QFRZN;
1335 				}
1336 				xpt_action(done_ccb);
1337 				goto out;
1338 			}
1339 		}
1340 		if (cam_periph_error(done_ccb,
1341 		    0, SF_RETRY_UA | SF_NO_PRINT) == ERESTART)
1342 			goto out;
1343 		if (done_ccb->ccb_h.status & CAM_DEV_QFRZN) {
1344 			cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1345 			done_ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1346 		}
1347 	} else {
1348 		/*
1349 		 * If we have successfully taken a device from the not
1350 		 * ready to ready state, re-scan the device and re-get
1351 		 * the inquiry information.  Many devices (mostly disks)
1352 		 * don't properly report their inquiry information unless
1353 		 * they are spun up.
1354 		 */
1355 		if (scsi_cmd->opcode == START_STOP_UNIT)
1356 			xpt_async(AC_INQ_CHANGED, done_ccb->ccb_h.path, NULL);
1357 	}
1358 
1359 	/*
1360 	 * Perform the final retry with the original CCB so that final
1361 	 * error processing is performed by the owner of the CCB.
1362 	 */
1363 	saved_ccb = (union ccb *)done_ccb->ccb_h.saved_ccb_ptr;
1364 	bcopy(saved_ccb, done_ccb, sizeof(*done_ccb));
1365 	xpt_free_ccb(saved_ccb);
1366 	if (done_ccb->ccb_h.cbfcnp != camperiphdone)
1367 		periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1368 	xpt_action(done_ccb);
1369 
1370 out:
1371 	/* Drop freeze taken due to CAM_DEV_QFREEZE flag set. */
1372 	cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1373 }
1374 
1375 /*
1376  * Generic Async Event handler.  Peripheral drivers usually
1377  * filter out the events that require personal attention,
1378  * and leave the rest to this function.
1379  */
1380 void
1381 cam_periph_async(struct cam_periph *periph, u_int32_t code,
1382 		 struct cam_path *path, void *arg)
1383 {
1384 	switch (code) {
1385 	case AC_LOST_DEVICE:
1386 		cam_periph_invalidate(periph);
1387 		break;
1388 	default:
1389 		break;
1390 	}
1391 }
1392 
1393 void
1394 cam_periph_bus_settle(struct cam_periph *periph, u_int bus_settle)
1395 {
1396 	struct ccb_getdevstats cgds;
1397 
1398 	xpt_setup_ccb(&cgds.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
1399 	cgds.ccb_h.func_code = XPT_GDEV_STATS;
1400 	xpt_action((union ccb *)&cgds);
1401 	cam_periph_freeze_after_event(periph, &cgds.last_reset, bus_settle);
1402 }
1403 
1404 void
1405 cam_periph_freeze_after_event(struct cam_periph *periph,
1406 			      struct timeval* event_time, u_int duration_ms)
1407 {
1408 	struct timeval delta;
1409 	struct timeval duration_tv;
1410 
1411 	if (!timevalisset(event_time))
1412 		return;
1413 
1414 	microtime(&delta);
1415 	timevalsub(&delta, event_time);
1416 	duration_tv.tv_sec = duration_ms / 1000;
1417 	duration_tv.tv_usec = (duration_ms % 1000) * 1000;
1418 	if (timevalcmp(&delta, &duration_tv, <)) {
1419 		timevalsub(&duration_tv, &delta);
1420 
1421 		duration_ms = duration_tv.tv_sec * 1000;
1422 		duration_ms += duration_tv.tv_usec / 1000;
1423 		cam_freeze_devq(periph->path);
1424 		cam_release_devq(periph->path,
1425 				RELSIM_RELEASE_AFTER_TIMEOUT,
1426 				/*reduction*/0,
1427 				/*timeout*/duration_ms,
1428 				/*getcount_only*/0);
1429 	}
1430 
1431 }
1432 
1433 static int
1434 camperiphscsistatuserror(union ccb *ccb, union ccb **orig_ccb,
1435     cam_flags camflags, u_int32_t sense_flags,
1436     int *openings, u_int32_t *relsim_flags,
1437     u_int32_t *timeout, u_int32_t *action, const char **action_string)
1438 {
1439 	int error;
1440 
1441 	switch (ccb->csio.scsi_status) {
1442 	case SCSI_STATUS_OK:
1443 	case SCSI_STATUS_COND_MET:
1444 	case SCSI_STATUS_INTERMED:
1445 	case SCSI_STATUS_INTERMED_COND_MET:
1446 		error = 0;
1447 		break;
1448 	case SCSI_STATUS_CMD_TERMINATED:
1449 	case SCSI_STATUS_CHECK_COND:
1450 		error = camperiphscsisenseerror(ccb, orig_ccb,
1451 					        camflags,
1452 					        sense_flags,
1453 					        openings,
1454 					        relsim_flags,
1455 					        timeout,
1456 					        action,
1457 					        action_string);
1458 		break;
1459 	case SCSI_STATUS_QUEUE_FULL:
1460 	{
1461 		/* no decrement */
1462 		struct ccb_getdevstats cgds;
1463 
1464 		/*
1465 		 * First off, find out what the current
1466 		 * transaction counts are.
1467 		 */
1468 		xpt_setup_ccb(&cgds.ccb_h,
1469 			      ccb->ccb_h.path,
1470 			      CAM_PRIORITY_NORMAL);
1471 		cgds.ccb_h.func_code = XPT_GDEV_STATS;
1472 		xpt_action((union ccb *)&cgds);
1473 
1474 		/*
1475 		 * If we were the only transaction active, treat
1476 		 * the QUEUE FULL as if it were a BUSY condition.
1477 		 */
1478 		if (cgds.dev_active != 0) {
1479 			int total_openings;
1480 
1481 			/*
1482 		 	 * Reduce the number of openings to
1483 			 * be 1 less than the amount it took
1484 			 * to get a queue full bounded by the
1485 			 * minimum allowed tag count for this
1486 			 * device.
1487 		 	 */
1488 			total_openings = cgds.dev_active + cgds.dev_openings;
1489 			*openings = cgds.dev_active;
1490 			if (*openings < cgds.mintags)
1491 				*openings = cgds.mintags;
1492 			if (*openings < total_openings)
1493 				*relsim_flags = RELSIM_ADJUST_OPENINGS;
1494 			else {
1495 				/*
1496 				 * Some devices report queue full for
1497 				 * temporary resource shortages.  For
1498 				 * this reason, we allow a minimum
1499 				 * tag count to be entered via a
1500 				 * quirk entry to prevent the queue
1501 				 * count on these devices from falling
1502 				 * to a pessimisticly low value.  We
1503 				 * still wait for the next successful
1504 				 * completion, however, before queueing
1505 				 * more transactions to the device.
1506 				 */
1507 				*relsim_flags = RELSIM_RELEASE_AFTER_CMDCMPLT;
1508 			}
1509 			*timeout = 0;
1510 			error = ERESTART;
1511 			*action &= ~SSQ_PRINT_SENSE;
1512 			break;
1513 		}
1514 		/* FALLTHROUGH */
1515 	}
1516 	case SCSI_STATUS_BUSY:
1517 		/*
1518 		 * Restart the queue after either another
1519 		 * command completes or a 1 second timeout.
1520 		 */
1521 		if ((sense_flags & SF_RETRY_BUSY) != 0 ||
1522 		    (ccb->ccb_h.retry_count--) > 0) {
1523 			error = ERESTART;
1524 			*relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT
1525 				      | RELSIM_RELEASE_AFTER_CMDCMPLT;
1526 			*timeout = 1000;
1527 		} else {
1528 			error = EIO;
1529 		}
1530 		break;
1531 	case SCSI_STATUS_RESERV_CONFLICT:
1532 	default:
1533 		error = EIO;
1534 		break;
1535 	}
1536 	return (error);
1537 }
1538 
1539 static int
1540 camperiphscsisenseerror(union ccb *ccb, union ccb **orig,
1541     cam_flags camflags, u_int32_t sense_flags,
1542     int *openings, u_int32_t *relsim_flags,
1543     u_int32_t *timeout, u_int32_t *action, const char **action_string)
1544 {
1545 	struct cam_periph *periph;
1546 	union ccb *orig_ccb = ccb;
1547 	int error, recoveryccb;
1548 
1549 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1550 	if (ccb->ccb_h.func_code == XPT_SCSI_IO && ccb->csio.bio != NULL)
1551 		biotrack(ccb->csio.bio, __func__);
1552 #endif
1553 
1554 	periph = xpt_path_periph(ccb->ccb_h.path);
1555 	recoveryccb = (ccb->ccb_h.cbfcnp == camperiphdone);
1556 	if ((periph->flags & CAM_PERIPH_RECOVERY_INPROG) && !recoveryccb) {
1557 		/*
1558 		 * If error recovery is already in progress, don't attempt
1559 		 * to process this error, but requeue it unconditionally
1560 		 * and attempt to process it once error recovery has
1561 		 * completed.  This failed command is probably related to
1562 		 * the error that caused the currently active error recovery
1563 		 * action so our  current recovery efforts should also
1564 		 * address this command.  Be aware that the error recovery
1565 		 * code assumes that only one recovery action is in progress
1566 		 * on a particular peripheral instance at any given time
1567 		 * (e.g. only one saved CCB for error recovery) so it is
1568 		 * imperitive that we don't violate this assumption.
1569 		 */
1570 		error = ERESTART;
1571 		*action &= ~SSQ_PRINT_SENSE;
1572 	} else {
1573 		scsi_sense_action err_action;
1574 		struct ccb_getdev cgd;
1575 
1576 		/*
1577 		 * Grab the inquiry data for this device.
1578 		 */
1579 		xpt_setup_ccb(&cgd.ccb_h, ccb->ccb_h.path, CAM_PRIORITY_NORMAL);
1580 		cgd.ccb_h.func_code = XPT_GDEV_TYPE;
1581 		xpt_action((union ccb *)&cgd);
1582 
1583 		err_action = scsi_error_action(&ccb->csio, &cgd.inq_data,
1584 		    sense_flags);
1585 		error = err_action & SS_ERRMASK;
1586 
1587 		/*
1588 		 * Do not autostart sequential access devices
1589 		 * to avoid unexpected tape loading.
1590 		 */
1591 		if ((err_action & SS_MASK) == SS_START &&
1592 		    SID_TYPE(&cgd.inq_data) == T_SEQUENTIAL) {
1593 			*action_string = "Will not autostart a "
1594 			    "sequential access device";
1595 			goto sense_error_done;
1596 		}
1597 
1598 		/*
1599 		 * Avoid recovery recursion if recovery action is the same.
1600 		 */
1601 		if ((err_action & SS_MASK) >= SS_START && recoveryccb) {
1602 			if (((err_action & SS_MASK) == SS_START &&
1603 			     ccb->csio.cdb_io.cdb_bytes[0] == START_STOP_UNIT) ||
1604 			    ((err_action & SS_MASK) == SS_TUR &&
1605 			     (ccb->csio.cdb_io.cdb_bytes[0] == TEST_UNIT_READY))) {
1606 				err_action = SS_RETRY|SSQ_DECREMENT_COUNT|EIO;
1607 				*relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1608 				*timeout = 500;
1609 			}
1610 		}
1611 
1612 		/*
1613 		 * If the recovery action will consume a retry,
1614 		 * make sure we actually have retries available.
1615 		 */
1616 		if ((err_action & SSQ_DECREMENT_COUNT) != 0) {
1617 		 	if (ccb->ccb_h.retry_count > 0 &&
1618 			    (periph->flags & CAM_PERIPH_INVALID) == 0)
1619 		 		ccb->ccb_h.retry_count--;
1620 			else {
1621 				*action_string = "Retries exhausted";
1622 				goto sense_error_done;
1623 			}
1624 		}
1625 
1626 		if ((err_action & SS_MASK) >= SS_START) {
1627 			/*
1628 			 * Do common portions of commands that
1629 			 * use recovery CCBs.
1630 			 */
1631 			orig_ccb = xpt_alloc_ccb_nowait();
1632 			if (orig_ccb == NULL) {
1633 				*action_string = "Can't allocate recovery CCB";
1634 				goto sense_error_done;
1635 			}
1636 			/*
1637 			 * Clear freeze flag for original request here, as
1638 			 * this freeze will be dropped as part of ERESTART.
1639 			 */
1640 			ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1641 			bcopy(ccb, orig_ccb, sizeof(*orig_ccb));
1642 		}
1643 
1644 		switch (err_action & SS_MASK) {
1645 		case SS_NOP:
1646 			*action_string = "No recovery action needed";
1647 			error = 0;
1648 			break;
1649 		case SS_RETRY:
1650 			*action_string = "Retrying command (per sense data)";
1651 			error = ERESTART;
1652 			break;
1653 		case SS_FAIL:
1654 			*action_string = "Unretryable error";
1655 			break;
1656 		case SS_START:
1657 		{
1658 			int le;
1659 
1660 			/*
1661 			 * Send a start unit command to the device, and
1662 			 * then retry the command.
1663 			 */
1664 			*action_string = "Attempting to start unit";
1665 			periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1666 
1667 			/*
1668 			 * Check for removable media and set
1669 			 * load/eject flag appropriately.
1670 			 */
1671 			if (SID_IS_REMOVABLE(&cgd.inq_data))
1672 				le = TRUE;
1673 			else
1674 				le = FALSE;
1675 
1676 			scsi_start_stop(&ccb->csio,
1677 					/*retries*/1,
1678 					camperiphdone,
1679 					MSG_SIMPLE_Q_TAG,
1680 					/*start*/TRUE,
1681 					/*load/eject*/le,
1682 					/*immediate*/FALSE,
1683 					SSD_FULL_SIZE,
1684 					/*timeout*/50000);
1685 			break;
1686 		}
1687 		case SS_TUR:
1688 		{
1689 			/*
1690 			 * Send a Test Unit Ready to the device.
1691 			 * If the 'many' flag is set, we send 120
1692 			 * test unit ready commands, one every half
1693 			 * second.  Otherwise, we just send one TUR.
1694 			 * We only want to do this if the retry
1695 			 * count has not been exhausted.
1696 			 */
1697 			int retries;
1698 
1699 			if ((err_action & SSQ_MANY) != 0) {
1700 				*action_string = "Polling device for readiness";
1701 				retries = 120;
1702 			} else {
1703 				*action_string = "Testing device for readiness";
1704 				retries = 1;
1705 			}
1706 			periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1707 			scsi_test_unit_ready(&ccb->csio,
1708 					     retries,
1709 					     camperiphdone,
1710 					     MSG_SIMPLE_Q_TAG,
1711 					     SSD_FULL_SIZE,
1712 					     /*timeout*/5000);
1713 
1714 			/*
1715 			 * Accomplish our 500ms delay by deferring
1716 			 * the release of our device queue appropriately.
1717 			 */
1718 			*relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1719 			*timeout = 500;
1720 			break;
1721 		}
1722 		default:
1723 			panic("Unhandled error action %x", err_action);
1724 		}
1725 
1726 		if ((err_action & SS_MASK) >= SS_START) {
1727 			/*
1728 			 * Drop the priority, so that the recovery
1729 			 * CCB is the first to execute.  Freeze the queue
1730 			 * after this command is sent so that we can
1731 			 * restore the old csio and have it queued in
1732 			 * the proper order before we release normal
1733 			 * transactions to the device.
1734 			 */
1735 			ccb->ccb_h.pinfo.priority--;
1736 			ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1737 			ccb->ccb_h.saved_ccb_ptr = orig_ccb;
1738 			error = ERESTART;
1739 			*orig = orig_ccb;
1740 		}
1741 
1742 sense_error_done:
1743 		*action = err_action;
1744 	}
1745 	return (error);
1746 }
1747 
1748 /*
1749  * Generic error handler.  Peripheral drivers usually filter
1750  * out the errors that they handle in a unique manner, then
1751  * call this function.
1752  */
1753 int
1754 cam_periph_error(union ccb *ccb, cam_flags camflags,
1755 		 u_int32_t sense_flags)
1756 {
1757 	struct cam_path *newpath;
1758 	union ccb  *orig_ccb, *scan_ccb;
1759 	struct cam_periph *periph;
1760 	const char *action_string;
1761 	cam_status  status;
1762 	int	    frozen, error, openings, devctl_err;
1763 	u_int32_t   action, relsim_flags, timeout;
1764 
1765 	action = SSQ_PRINT_SENSE;
1766 	periph = xpt_path_periph(ccb->ccb_h.path);
1767 	action_string = NULL;
1768 	status = ccb->ccb_h.status;
1769 	frozen = (status & CAM_DEV_QFRZN) != 0;
1770 	status &= CAM_STATUS_MASK;
1771 	devctl_err = openings = relsim_flags = timeout = 0;
1772 	orig_ccb = ccb;
1773 
1774 	/* Filter the errors that should be reported via devctl */
1775 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1776 	case CAM_CMD_TIMEOUT:
1777 	case CAM_REQ_ABORTED:
1778 	case CAM_REQ_CMP_ERR:
1779 	case CAM_REQ_TERMIO:
1780 	case CAM_UNREC_HBA_ERROR:
1781 	case CAM_DATA_RUN_ERR:
1782 	case CAM_SCSI_STATUS_ERROR:
1783 	case CAM_ATA_STATUS_ERROR:
1784 	case CAM_SMP_STATUS_ERROR:
1785 		devctl_err++;
1786 		break;
1787 	default:
1788 		break;
1789 	}
1790 
1791 	switch (status) {
1792 	case CAM_REQ_CMP:
1793 		error = 0;
1794 		action &= ~SSQ_PRINT_SENSE;
1795 		break;
1796 	case CAM_SCSI_STATUS_ERROR:
1797 		error = camperiphscsistatuserror(ccb, &orig_ccb,
1798 		    camflags, sense_flags, &openings, &relsim_flags,
1799 		    &timeout, &action, &action_string);
1800 		break;
1801 	case CAM_AUTOSENSE_FAIL:
1802 		error = EIO;	/* we have to kill the command */
1803 		break;
1804 	case CAM_UA_ABORT:
1805 	case CAM_UA_TERMIO:
1806 	case CAM_MSG_REJECT_REC:
1807 		/* XXX Don't know that these are correct */
1808 		error = EIO;
1809 		break;
1810 	case CAM_SEL_TIMEOUT:
1811 		if ((camflags & CAM_RETRY_SELTO) != 0) {
1812 			if (ccb->ccb_h.retry_count > 0 &&
1813 			    (periph->flags & CAM_PERIPH_INVALID) == 0) {
1814 				ccb->ccb_h.retry_count--;
1815 				error = ERESTART;
1816 
1817 				/*
1818 				 * Wait a bit to give the device
1819 				 * time to recover before we try again.
1820 				 */
1821 				relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1822 				timeout = periph_selto_delay;
1823 				break;
1824 			}
1825 			action_string = "Retries exhausted";
1826 		}
1827 		/* FALLTHROUGH */
1828 	case CAM_DEV_NOT_THERE:
1829 		error = ENXIO;
1830 		action = SSQ_LOST;
1831 		break;
1832 	case CAM_REQ_INVALID:
1833 	case CAM_PATH_INVALID:
1834 	case CAM_NO_HBA:
1835 	case CAM_PROVIDE_FAIL:
1836 	case CAM_REQ_TOO_BIG:
1837 	case CAM_LUN_INVALID:
1838 	case CAM_TID_INVALID:
1839 	case CAM_FUNC_NOTAVAIL:
1840 		error = EINVAL;
1841 		break;
1842 	case CAM_SCSI_BUS_RESET:
1843 	case CAM_BDR_SENT:
1844 		/*
1845 		 * Commands that repeatedly timeout and cause these
1846 		 * kinds of error recovery actions, should return
1847 		 * CAM_CMD_TIMEOUT, which allows us to safely assume
1848 		 * that this command was an innocent bystander to
1849 		 * these events and should be unconditionally
1850 		 * retried.
1851 		 */
1852 	case CAM_REQUEUE_REQ:
1853 		/* Unconditional requeue if device is still there */
1854 		if (periph->flags & CAM_PERIPH_INVALID) {
1855 			action_string = "Periph was invalidated";
1856 			error = EIO;
1857 		} else if (sense_flags & SF_NO_RETRY) {
1858 			error = EIO;
1859 			action_string = "Retry was blocked";
1860 		} else {
1861 			error = ERESTART;
1862 			action &= ~SSQ_PRINT_SENSE;
1863 		}
1864 		break;
1865 	case CAM_RESRC_UNAVAIL:
1866 		/* Wait a bit for the resource shortage to abate. */
1867 		timeout = periph_noresrc_delay;
1868 		/* FALLTHROUGH */
1869 	case CAM_BUSY:
1870 		if (timeout == 0) {
1871 			/* Wait a bit for the busy condition to abate. */
1872 			timeout = periph_busy_delay;
1873 		}
1874 		relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1875 		/* FALLTHROUGH */
1876 	case CAM_ATA_STATUS_ERROR:
1877 	case CAM_REQ_CMP_ERR:
1878 	case CAM_CMD_TIMEOUT:
1879 	case CAM_UNEXP_BUSFREE:
1880 	case CAM_UNCOR_PARITY:
1881 	case CAM_DATA_RUN_ERR:
1882 	default:
1883 		if (periph->flags & CAM_PERIPH_INVALID) {
1884 			error = EIO;
1885 			action_string = "Periph was invalidated";
1886 		} else if (ccb->ccb_h.retry_count == 0) {
1887 			error = EIO;
1888 			action_string = "Retries exhausted";
1889 		} else if (sense_flags & SF_NO_RETRY) {
1890 			error = EIO;
1891 			action_string = "Retry was blocked";
1892 		} else {
1893 			ccb->ccb_h.retry_count--;
1894 			error = ERESTART;
1895 		}
1896 		break;
1897 	}
1898 
1899 	if ((sense_flags & SF_PRINT_ALWAYS) ||
1900 	    CAM_DEBUGGED(ccb->ccb_h.path, CAM_DEBUG_INFO))
1901 		action |= SSQ_PRINT_SENSE;
1902 	else if (sense_flags & SF_NO_PRINT)
1903 		action &= ~SSQ_PRINT_SENSE;
1904 	if ((action & SSQ_PRINT_SENSE) != 0)
1905 		cam_error_print(orig_ccb, CAM_ESF_ALL, CAM_EPF_ALL);
1906 	if (error != 0 && (action & SSQ_PRINT_SENSE) != 0) {
1907 		if (error != ERESTART) {
1908 			if (action_string == NULL)
1909 				action_string = "Unretryable error";
1910 			xpt_print(ccb->ccb_h.path, "Error %d, %s\n",
1911 			    error, action_string);
1912 		} else if (action_string != NULL)
1913 			xpt_print(ccb->ccb_h.path, "%s\n", action_string);
1914 		else {
1915 			xpt_print(ccb->ccb_h.path,
1916 			    "Retrying command, %d more tries remain\n",
1917 			    ccb->ccb_h.retry_count);
1918 		}
1919 	}
1920 
1921 	if (devctl_err && (error != 0 || (action & SSQ_PRINT_SENSE) != 0))
1922 		cam_periph_devctl_notify(orig_ccb);
1923 
1924 	if ((action & SSQ_LOST) != 0) {
1925 		lun_id_t lun_id;
1926 
1927 		/*
1928 		 * For a selection timeout, we consider all of the LUNs on
1929 		 * the target to be gone.  If the status is CAM_DEV_NOT_THERE,
1930 		 * then we only get rid of the device(s) specified by the
1931 		 * path in the original CCB.
1932 		 */
1933 		if (status == CAM_SEL_TIMEOUT)
1934 			lun_id = CAM_LUN_WILDCARD;
1935 		else
1936 			lun_id = xpt_path_lun_id(ccb->ccb_h.path);
1937 
1938 		/* Should we do more if we can't create the path?? */
1939 		if (xpt_create_path(&newpath, periph,
1940 				    xpt_path_path_id(ccb->ccb_h.path),
1941 				    xpt_path_target_id(ccb->ccb_h.path),
1942 				    lun_id) == CAM_REQ_CMP) {
1943 
1944 			/*
1945 			 * Let peripheral drivers know that this
1946 			 * device has gone away.
1947 			 */
1948 			xpt_async(AC_LOST_DEVICE, newpath, NULL);
1949 			xpt_free_path(newpath);
1950 		}
1951 	}
1952 
1953 	/* Broadcast UNIT ATTENTIONs to all periphs. */
1954 	if ((action & SSQ_UA) != 0)
1955 		xpt_async(AC_UNIT_ATTENTION, orig_ccb->ccb_h.path, orig_ccb);
1956 
1957 	/* Rescan target on "Reported LUNs data has changed" */
1958 	if ((action & SSQ_RESCAN) != 0) {
1959 		if (xpt_create_path(&newpath, NULL,
1960 				    xpt_path_path_id(ccb->ccb_h.path),
1961 				    xpt_path_target_id(ccb->ccb_h.path),
1962 				    CAM_LUN_WILDCARD) == CAM_REQ_CMP) {
1963 
1964 			scan_ccb = xpt_alloc_ccb_nowait();
1965 			if (scan_ccb != NULL) {
1966 				scan_ccb->ccb_h.path = newpath;
1967 				scan_ccb->ccb_h.func_code = XPT_SCAN_TGT;
1968 				scan_ccb->crcn.flags = 0;
1969 				xpt_rescan(scan_ccb);
1970 			} else {
1971 				xpt_print(newpath,
1972 				    "Can't allocate CCB to rescan target\n");
1973 				xpt_free_path(newpath);
1974 			}
1975 		}
1976 	}
1977 
1978 	/* Attempt a retry */
1979 	if (error == ERESTART || error == 0) {
1980 		if (frozen != 0)
1981 			ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1982 		if (error == ERESTART)
1983 			xpt_action(ccb);
1984 		if (frozen != 0)
1985 			cam_release_devq(ccb->ccb_h.path,
1986 					 relsim_flags,
1987 					 openings,
1988 					 timeout,
1989 					 /*getcount_only*/0);
1990 	}
1991 
1992 	return (error);
1993 }
1994 
1995 #define CAM_PERIPH_DEVD_MSG_SIZE	256
1996 
1997 static void
1998 cam_periph_devctl_notify(union ccb *ccb)
1999 {
2000 	struct cam_periph *periph;
2001 	struct ccb_getdev *cgd;
2002 	struct sbuf sb;
2003 	int serr, sk, asc, ascq;
2004 	char *sbmsg, *type;
2005 
2006 	sbmsg = malloc(CAM_PERIPH_DEVD_MSG_SIZE, M_CAMPERIPH, M_NOWAIT);
2007 	if (sbmsg == NULL)
2008 		return;
2009 
2010 	sbuf_new(&sb, sbmsg, CAM_PERIPH_DEVD_MSG_SIZE, SBUF_FIXEDLEN);
2011 
2012 	periph = xpt_path_periph(ccb->ccb_h.path);
2013 	sbuf_printf(&sb, "device=%s%d ", periph->periph_name,
2014 	    periph->unit_number);
2015 
2016 	sbuf_printf(&sb, "serial=\"");
2017 	if ((cgd = (struct ccb_getdev *)xpt_alloc_ccb_nowait()) != NULL) {
2018 		xpt_setup_ccb(&cgd->ccb_h, ccb->ccb_h.path,
2019 		    CAM_PRIORITY_NORMAL);
2020 		cgd->ccb_h.func_code = XPT_GDEV_TYPE;
2021 		xpt_action((union ccb *)cgd);
2022 
2023 		if (cgd->ccb_h.status == CAM_REQ_CMP)
2024 			sbuf_bcat(&sb, cgd->serial_num, cgd->serial_num_len);
2025 		xpt_free_ccb((union ccb *)cgd);
2026 	}
2027 	sbuf_printf(&sb, "\" ");
2028 	sbuf_printf(&sb, "cam_status=\"0x%x\" ", ccb->ccb_h.status);
2029 
2030 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
2031 	case CAM_CMD_TIMEOUT:
2032 		sbuf_printf(&sb, "timeout=%d ", ccb->ccb_h.timeout);
2033 		type = "timeout";
2034 		break;
2035 	case CAM_SCSI_STATUS_ERROR:
2036 		sbuf_printf(&sb, "scsi_status=%d ", ccb->csio.scsi_status);
2037 		if (scsi_extract_sense_ccb(ccb, &serr, &sk, &asc, &ascq))
2038 			sbuf_printf(&sb, "scsi_sense=\"%02x %02x %02x %02x\" ",
2039 			    serr, sk, asc, ascq);
2040 		type = "error";
2041 		break;
2042 	case CAM_ATA_STATUS_ERROR:
2043 		sbuf_printf(&sb, "RES=\"");
2044 		ata_res_sbuf(&ccb->ataio.res, &sb);
2045 		sbuf_printf(&sb, "\" ");
2046 		type = "error";
2047 		break;
2048 	default:
2049 		type = "error";
2050 		break;
2051 	}
2052 
2053 	if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
2054 		sbuf_printf(&sb, "CDB=\"");
2055 		scsi_cdb_sbuf(scsiio_cdb_ptr(&ccb->csio), &sb);
2056 		sbuf_printf(&sb, "\" ");
2057 	} else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
2058 		sbuf_printf(&sb, "ACB=\"");
2059 		ata_cmd_sbuf(&ccb->ataio.cmd, &sb);
2060 		sbuf_printf(&sb, "\" ");
2061 	}
2062 
2063 	if (sbuf_finish(&sb) == 0)
2064 		devctl_notify("CAM", "periph", type, sbuf_data(&sb));
2065 	sbuf_delete(&sb);
2066 	free(sbmsg, M_CAMPERIPH);
2067 }
2068 
2069